<div class="bt jv af dn dj dk">Over the last decade or so, the world of AI has been obsessed with the problem of pattern recognition &mdash; Ask any budding engineer &ldquo;Give me an algorithm to find&nbsp;X in an environment&nbsp;Y&rdquo;, and expect to hear the words, &ldquo;Give me&nbsp;T examples in environment&nbsp;Y, and I will predict&nbsp;X&nbsp;with so much accuracy&rdquo;.The recent advancement in ML techniques have equipped upcoming entrepreneurs with the belief that computers can do magic.</div><div class="fj"><div class="v w"><div class="lc ld le lf lg lh ab li ac lj af"><figure class="ll lm ln lo lp fj"><div class="ly l et"> </div><div class="ly l et"> <iframe src="https://www.youtube.com/embed/VOC3huqHrss?&wmode=opaque" frameborder="0" allowfullscreen="" class="fr-draggable" style="width: 687px; height: 377px;"></iframe></div></figure></div></div></div>Having watched this video makes me think who are we to say otherwise. To add to this, a recent stroll at the CVPR&rsquo;17 poster session was convincing enough that the academic world is testing the strength of these techniques by solving previously unimaginable and jaw-dropping problems. These algorithms are passing tests with flying colors.A basic computer is actually a pretty dumb entity working on binary electrical voltages and so it is safe to assume it can do no magic, and thus is heavily reliant on smart algorithms and tuning to solve the aforementioned problems.This brings us to an important question,<h1 class="mu mv ef aq ap ed mw mx my mz na nb nc nd ne nf ng" data-selectable-paragraph="" id="1883">How can I make a computer solve my problems?</h1>The answer to this question has seen many paradigm shifts since computers were created, and the current approach is exactly what our budding engineer had mentioned: build &lsquo;smart algorithms&rsquo; and present the computer with &lsquo;enough&rsquo; real-world examples of the environment (training data), so that when the computer sees &lsquo;similar data&rsquo;, it knows what to do. With the business of &ldquo;smart algorithms&rdquo; been taken care of by computer scientists all over the world, one piece of the puzzle that might still haunt upcoming startups is &ldquo;how much data is actually enough?&rdquo;.Before answering this question, let&rsquo;s try to understand why is this even a question. Why can&rsquo;t we have infinite data and just focus on solving the problem? The reason is every data collected has a $ value associated with it, since data is usually collected manually, and so people have to be trained and compensated for the&nbsp;<a class="cz bv nm nn no np" href="https://playment.io/data-collection" rel="noopener" target="_blank">tedious work of data collection</a>, and so if I can do with&nbsp;Ttraining samples, I will not collect 10&nbsp;X T training samples. This brings us back to the main question: How much is this&nbsp;T for my problem, and can I do with lesser&nbsp;T? The answer to this question depends mainly on two factors:<ul><li class="me mf ef aq mg b mh mi mj mk ml mm mn mo mp mq mr nq nr ns" data-selectable-paragraph="" id="ed9e">How catastrophic is it for my product to make errors?</li><li class="me mf ef aq mg b mh nt mj nu ml nv mn nw mp nx mr nq nr ns" data-selectable-paragraph="" id="b6d6">How diverse can be the input to my product?</li></ul>While everyone wants to come up with a perfect product, real-world ML products are seldom perfect. They always have a potential to make errors with different products having different tolerances for error. For example, an application that predicts which team will win a football game has a higher tolerance than the goal-line technology to assess whether a goal was scored.Whatever be the product type, no one likes products which make mistakes, and if your business revolves around customer satisfaction, 99.0 % accuracy is more important than 98.95 % accuracy, and so every new training point matters. Imagine a 0.0001 % accuracy reduction in the left/right prediction in your&nbsp;<a class="cz bv nm nn no np" href="https://playment.io/adas/?utm_source=hackernoon&utm_content=how_much_training_data" rel="noopener" target="_blank">self-driving car</a>, and you would never think about reducing your training data.Just like companies want error-free products, they also want products that work in any environment in this world. &ldquo;A face detection algorithm developed for detecting faces in Africa should also detect an alien&rsquo;s face when they arrive in 2050&rdquo;. To achieve such high standards, the algorithm needs to be presented with every input possibility, so it learns to better the notion of &ldquo;similar data&rdquo;.The recent growth of deep learning architectures is necessary because our world is complex. To build something reliable to handle all its subtleties, we need to train our product as much as we can. To illustrate the impact of deep architectures on training data size, let us analyze the problem of face detection/recognition.The famous Haar Classifier or SVM based face detectors can be trained on a few thousand samples to obtain their optimal performance. In comparison to this, the&nbsp;<a class="cz bv nm nn no np" href="https://github.com/davidsandberg/facenet" rel="noopener" target="_blank">FaceNet</a> architecture for detecting and recognizing faces was trained on more than 400K training data samples, and it has effectively &ldquo;solved&rdquo; the closed set face recognition problem. However, detecting faces is a relatively simple problem in the field of Computer Vision, and more complex problems of ML (language translation, self-driving data) would only need more and more data before they are &ldquo;solved&rdquo;. If only Apple had access to infinite training, SIRI would have been able to recognize my Indian accent as well as it recognizes the accent of my American colleagues.The cost of collecting training data has inspired scientists to look for alternative approaches for manual data annotations. Two approaches at the forefront are, transfer learning and synthetic data generation. Using a lot of data to train a dog detector and using this as a starting point to detect cats is a simple example of transfer learning. An important aspect to understand while discussing transfer is that it can utilize the powers of unsupervised learning, which means training data without annotations (which is extremely easy to collect). Algorithms can be built to learn important features from these unlabelled data, and eventually tuned with limited training data. Synthetic data can be considered heuristics-based computer-generated data to best simulate the environment. The importance of this direction is substantiated by&nbsp;<a class="cz bv nm nn no np" href="https://machinelearning.apple.com/2017/07/07/GAN.html" rel="noopener" target="_blank">Apple&rsquo;s recent effort</a>, which received the best paper award in CVPR 2017. These approaches are still in their early stages of development. Only time will tell whether we can replace&nbsp;<a class="cz bv nm nn no np" href="https://blog.playment.io/training-data-for-computer-vision/utm_source=hackernoon&utm_medium=referral&utm_content=how_much_training_data" rel="noopener" target="_blank">human annotations</a> altogether.Two inevitable aspects which increase the necessity of collecting more training data are human error and data shelf-life. Even though modern ML algorithms are robust to noise in training data, human errors end up reducing the effective training size in almost all cases. To add to this, data collected using sensors (camera, microphone) suffer from the fact that what is state-of-the-art data collection now might be insufficient in a few years &mdash; Imagine classifying the ImageNet images using pixelated images obtained 5 years ago.The quantity and quality of training data is undeniably a hot topic right now people are exploring different techniques to solve this issue. In a few years, we might get a more clear picture of how much&nbsp;<a class="cz bv nm nn no np" href="https://playment.io/utm_source=hackernoon&utm_medium=referral&utm_content=how_much_training_data" rel="noopener" target="_blank">training data</a> is sufficient for a problem. But for now it would be safe to conclude&nbsp;&ldquo;the more, the better&rdquo;.

Over the last decade or so, the world of AI has been obsessed with the problem of pattern recognition — Ask any budding engineer “Give me an algorithm to find X in an environment Y”, and expect to hear the words, “Give me T examples in environment Y, and I will predict X with so much accuracy”.

Why Does Machine Learning Require So Much Training Data?

Henrik Christensen shares his views on the impact of the robot revolution and stories from his life in European and US robotics research.

The impact of the robot revolution with roboticist Henrik Christensen

Even though it has the right of way, the connected and automated Lincoln MKZ stops at an intersection because another vehicle is running the light. Vehicle-to-vehicle communication enabled this safety feature. Photo by Michigan Photography

Mcity demonstrates the powerful pairing of connectivity and automation.

How self-driving cars can be safer with connected technology

The first programmable memristor computer&mdash;not just a memristor array operated through an external computer&mdash;has been developed at the University of Michigan. It could lead to AI processing directly on small, energy-constrained devices such as smartphones and sensors.A smartphone AI processor would mean that voice commands would no longer have to be sent to the cloud for interpretation, speeding up response time.&ldquo;Everyone wants to put an AI processor on smartphones, but you don&rsquo;t want your cell phone battery to drain very quickly,&rdquo; said <a href="http://www-personal.umich.edu/~wluee/" target="_blank">Wei Lu</a>, a professor of electrical and computer engineering at U-M and senior author of the paper in Nature Electronics.In medical devices, the ability to run AI algorithms without the cloud would enable better security and privacy.<h4>Why memristors are good for machine learning</h4>The key to making this possible could be an advanced computer component called the memristor. This circuit element, an electrical resistor with a memory, has a variable resistance that can serve as a form of information storage. Because memristors store and process information in the same location, they can get around the biggest bottleneck for computing speed and power: the connection between memory and processor.This is especially important for machine-learning algorithms that deal with lots of data &nbsp;to do things like identify objects in photos and videos&mdash;or predict which hospital patients are at higher risk of infection. Already, programmers prefer to run these algorithms on graphical processing units (GPUs) rather than a computer&rsquo;s main processor, the central processing unit (CPU).&ldquo;GPUs and very customized and optimized digital circuits are considered to be about 10-100 times better than CPUs in terms of power and throughput. Memristor AI processors could be another 10-100 times better,&rdquo; said Lu.GPUs perform better at machine learning tasks because they have thousands of small cores for running calculations all at once, as opposed to the string of calculations waiting their turn on one of the few powerful cores in a CPU. A memristor array takes this even further. Each memristor is able to do its own calculation, allowing thousands of operations within a core to be performed at once. In this experimental-scale computer, there were more than 5,800 memristors. A commercial design could include millions of them.Memristor arrays are especially suited to machine learning problems. The reason for this is the way that machine learning algorithms turn data into vectors&mdash;essentially, lists of data points. In predicting a patient&rsquo;s risk of infection in a hospital, for instance, this vector might list numerical representations of a patient&rsquo;s risk factors. Then, machine learning algorithms compare these &ldquo;input&rdquo; vectors with &ldquo;feature&rdquo; vectors stored in memory.&nbsp;These feature vectors represent certain traits of the data (such as the presence of an underlying disease). If matched, the system knows that the input data has that trait. The vectors are stored in matrices, which are like the spreadsheets of mathematics, and these matrices can be mapped directly onto the memristor arrays.&nbsp;What&rsquo;s more, as data is fed through the array, the bulk of the mathematical processing occurs through the natural resistances in the memristors, eliminating the need to move feature vectors in and out of the memory to perform the computations. This makes the &nbsp;arrays highly efficient at complicated matrix calculations.&nbsp;Earlier studies demonstrated the potential of memristor arrays for speeding up machine learning, but they needed external computing elements to function.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1569831908719-memristors-9content.jpg" style="width: 594px;" class="fr-fic fr-dib">Wei Lu stands with Seung Hwan Lee, an electrical engineering PhD student and co-author, who holds the memristor array. Photo: Robert Coelius, Michigan Engineering Communications &amp; Marketing&nbsp;<h4>Building a programmable memristor computer</h4>To build the first programmable memristor computer, Lu&rsquo;s team worked with <a href="https://zhang.engin.umich.edu/" target="_blank">Zhengya Zhang</a> and <a href="https://www.mpflynngroup.com/michael-flynn.html" target="_blank">Michael Flynn</a>, an associate professor and a professor of electrical and computer engineering at U-M, to design a chip that could integrate the memristor array with all the other elements needed to program and run it. Those components included a conventional digital processor and communication channels, as well as digital/analog converters to serve as interpreters between the analog memristor array and the rest of the computer.&nbsp;Lu&rsquo;s team then integrated the memristor array directly on the chip at U-M&rsquo;s Lurie Nanofabrication Facility. They also developed software to map machine learning algorithms onto the matrix-like structure of the memristor array.The team demonstrated the device with three bread-and-butter machine learning algorithms:<ul><li>Perceptron, which is used to classify information. They were able to identify imperfect Greek letters with 100% accuracy.</li><li>Sparse coding, which compresses and categorizes data, particularly images. The computer was able to find the most efficient way to reconstruct images in a set and identified patterns with 100% accuracy.</li><li>Two-layer neural network, designed to find patterns in complex data. This two-layer network found commonalities and differentiating factors in breast cancer screening data and then classified each case as malignant or benign with 94.6% accuracy.</li></ul>There are challenges in scaling up for commercial use: Memristors can&rsquo;t yet be made as identical as they need to be, and the information stored in the array isn&rsquo;t entirely reliable because it runs on analog&rsquo;s continuum rather than the digital either/or. These are future directions of Lu&rsquo;s group.

The memristor array chip plugs into the custom computer chip, forming the first programmable memristor computer. The team demonstrated that it could run three standard types of machine learning algorithms. Photo: Robert Coelius, Michigan Engineering Communications & Marketing

Circuit elements that store information in their electrical resistances enable a brain-like form of computing, storing and processing information in the same place.

First programmable memristor computer aims to bring AI processing down from the cloud

Most recent advances in artificial-intelligence systems such as speech- or face-recognition programs have come courtesy of neural networks, densely interconnected meshes of simple information processors that learn to perform tasks by analyzing huge sets of training data.But neural nets are large, and their computations are energy intensive, so they&rsquo;re not very practical for handheld devices. Most smartphone apps that rely on neural nets simply upload data to internet servers, which process it and send the results back to the phone.Now, MIT researchers have developed a special-purpose chip that increases the speed of neural-network computations by three to seven times over its predecessors, while reducing power consumption 94 to 95 percent. That could make it practical to run neural networks locally on smartphones or even to embed them in household appliances.&ldquo;The general processor model is that there is a memory in some part of the chip, and there is a processor in another part of the chip, and you move the data back and forth between them when you do these computations,&rdquo; says Avishek Biswas, an MIT graduate student in electrical engineering and computer science, who led the new chip&rsquo;s development.&ldquo;Since these machine-learning algorithms need so many computations, this transferring back and forth of data is the dominant portion of the energy consumption. But the computation these algorithms do can be simplified to one specific operation, called the dot product. Our approach was, can we implement this dot-product functionality inside the memory so that you don&rsquo;t need to transfer this data back and forth?&rdquo;Biswas and his thesis advisor, Anantha Chandrakasan, dean of MIT&rsquo;s School of Engineering and the Vannevar Bush Professor of Electrical Engineering and Computer Science, describe the new chip in a paper that Biswas is presenting this week at the International Solid State Circuits Conference.Back to analogNeural networks are typically arranged into layers. A single processing node in one layer of the network will generally receive data from several nodes in the layer below and pass data to several nodes in the layer above. Each connection between nodes has its own &ldquo;weight,&rdquo; which indicates how large a role the output of one node will play in the computation performed by the next. Training the network is a matter of setting those weights.A node receiving data from multiple nodes in the layer below will multiply each input by the weight of the corresponding connection and sum the results. That operation &mdash; the summation of multiplications &mdash; is the definition of a dot product. If the dot product exceeds some threshold value, the node will transmit it to nodes in the next layer, over connections with their own weights.A neural net is an abstraction: The &ldquo;nodes&rdquo; are just weights stored in a computer&rsquo;s memory. Calculating a dot product usually involves fetching a weight from memory, fetching the associated data item, multiplying the two, storing the result somewhere, and then repeating the operation for every input to a node. Given that a neural net will have thousands or even millions of nodes, that&rsquo;s a lot of data to move around.But that sequence of operations is just a digital approximation of what happens in the brain, where signals traveling along multiple neurons meet at a &ldquo;synapse,&rdquo; or a gap between bundles of neurons. The neurons&rsquo; firing rates and the electrochemical signals that cross the synapse correspond to the data values and weights. The MIT researchers&rsquo; new chip improves efficiency by replicating the brain more faithfully.In the chip, a node&rsquo;s input values are converted into electrical voltages and then multiplied by the appropriate weights. Summing the products is simply a matter of combining the voltages. Only the combined voltages are converted back into a digital representation and stored for further processing.The chip can thus calculate dot products for multiple nodes &mdash; 16 at a time, in the prototype &mdash; in a single step, instead of shuttling between a processor and memory for every computation.All or nothingOne of the keys to the system is that all the weights are either 1 or -1. That means that they can be implemented within the memory itself as simple switches that either close a circuit or leave it open. Recent theoretical work suggests that neural nets trained with only two weights should lose little accuracy &mdash; somewhere between 1 and 2 percent.Biswas and Chandrakasan&rsquo;s research bears that prediction out. In experiments, they ran the full implementation of a neural network on a conventional computer and the binary-weight equivalent on their chip. Their chip&rsquo;s results were generally within 2 to 3 percent of the conventional network&rsquo;s.&quot;This is a promising real-world demonstration of SRAM-based in-memory analog computing for deep-learning applications,&rdquo; says Dario Gil, vice president of artificial intelligence at IBM. &quot;The results show impressive specifications for the energy-efficient implementation of convolution operations with memory arrays. It certainly will open the possibility to employ more complex convolutional neural networks for image and video classifications in IoT [the internet of things] in the future.&quot;

MIT researchers have developed a special-purpose chip that increases the speed of neural-network computations by three to seven times over its predecessors, while reducing power consumption 93 to 96 percent. That could make it practical to run neural networks locally on smartphones or even to embed them in household appliances.  Image: Chelsea Turner/MIT

New chip reduces neural networks’ power consumption by up to 95 percent, making them practical for battery-powered devices.

Neural networks everywhere

Most robots used in factories today are powerful but blind. They are programmed to do specific tasks repeatedly with high precision and speed, but until now have not been able to adapt to changes in the environment or handling variables. Robot arms often lack the ability to grasp an object if it is moved a few inches, not to mention picking or sorting millions of different products in a warehouse.As a result system integrators design shaker tables, feeders, conveyor belts, etc., so robots can be offered the same parts in the right orientation. Larger industrial robots need cages so they don’t accidentally hurt nearby people. Unlike human hands, which can perform tens of thousands of different tasks, robot arms can only be programmed for single tasks, and they need the proper end effectors.Additionally, most robots cannot sense, unless you equip them with force or vision sensors. They don’t know if tasks are performed properly, because they cannot see or feel.Traditional computer vision requires pre-registration of all objects, i.e. scanning and building computer-aided design (CAD) models for all items beforehand with expensive 3D cameras. The process is time-consuming and inflexible. Even if there was a box with exactly the same shape, but a different size, a robot couldn’t automatically recognize the differences and figure out how to grasp the new item.In addition, translucent packaging, reflective surfaces, and deformable objects (such as food and clothing) make robotic vision an even more difficult problem to solve. The additional non-robot hardware and integration end up costing four times more than a robot arm, which costs anywhere between $10,000 to $100,000.Adding intelligence to robotsTo solve these issues, deep reinforcement learning (DRL) is unleashing the full potential of robotics, giving them the ability to recognize and react to their surroundings and handle product variations. Robots can now learn from experiences with off-the-shelf cameras and deep neural networks.With enough data and practice, robots can now teach themselves new abilities, as was seen with robots’ ability to identify images and win video games. With every grasp and trial, a robot gets smarter and more adept at mastering the task. Furthermore, cloud-connected robots allow learnings to be shared with each other. This is a tremendous shift, making robotics solutions more dexterous, flexible, and scalable.Deep Learning (DL) is primarily used for image classification. Deep neural networks learn representations, rather than task-specific algorithms.Reinforcement learning (RL) is inspired by behavioral psychology. It learns control policies by trial-and-error through reward or punishment.Deep Reinforcement Learning (DRL) combines these two approaches. It is currently suited for:· Simple or simulated tasks· Fault-tolerant tasks· Just enough variation to make explicit modeling difficult· Easy-to-define goals (reward functions)· Semi-structured environmentsWarehouse as a testing ground for AIAutomation in warehouses is widely considered the “low-hanging fruit” for these technologies because tasks tend to be similar across warehouses, and the environment is more structured. For example, order picking represents more than 40% of operational costs across most warehouses. While tasks are repetitive, they involve some variation with products and packaging design, making it impossible to hard-program robots in advance.Furthermore, labor costs account for up to 70% of a warehouse’s total budget. Pressured by e-commerce companies such as Amazon, retailers are looking to automation to address labor shortages and reduce costs. Customer demand, price competition, and fast delivery requirements are driving a strong business case for warehouses to be the first adopter of AI-enabled robots.Chart: Why do we see AI-enabled robots in the warehouse first?<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNTY4OTAzMzQ5MzI3LTFfTDhkY3A1c1lfcEU0MTdqdW9VaEMwZy5wbmciLCJlZGl0cyI6eyJyZXNpemUiOnsid2lkdGgiOjk1MCwiZml0IjoiY292ZXIifX19" style="width: 766px;" class="fr-fic fr-dib"> Most tasks in a warehouse are fault tolerant, unlike tasks on a production floor or in the automotive industry. If a robot accidentally drops an object in a warehouse, it can try again as long as it achieves the required speed and success rates without damaging the goods.DRL is particularly suited for fault-tolerant simple tasks — actions a human can perform within a few seconds — with enough variation in semi-structured environments. As technology improves in environments like warehouses, it can then be applied in other industries, such as manufacturing, where tasks are more complicated and more uncertainties exist. Thus, an automated warehouse acts as an incubator, predicting a knowledge spill-over to many other market segments.Piece-picking is the holy grailWarehouse automation has been around for decades, in the form of automated storage, retrieval systems, and mobile robots like those from Kiva (now Amazon Robotics), that move items around.But most of the pick-and-pack operations within a warehouse are still performed by humans, with labor costs accounting for 50 to 70% of the overall warehouse budget. Worker shortages have driven up salaries by a rate of 6 to 8% every year, while robot costs have fallen considerably since the 1990s.Piece-picking has long been the holy grail of robotics. Year after year, incumbents such as Amazon and KUKA host robotics challenges for startups and academic teams to create machines that can identify, pick up and stow goods. The emergence of deep learning allows robots to recognize, pick, and place hundreds or up to thousands of items.But the technology is not yet perfect. It’s still challenging for a machine to recognize tens of millions of objects, and manipulate deformable objects or items with transparent packages. However, our interviews with Locus Robotics and Osaro indicate that many industry experts expect the technology will mature. In addition, we will have a robot vision camera that can recognize almost everything in a warehouse within five years. If that happens, it will affect not just warehouses, but retail, delivery, and many other applications that we haven’t yet imagined.Pushing towards standard modelsTechnology improvements in warehouse automation and beyond will also push the mature, but fragmented, robotics industry to become more standardized and modular. Given the nature of narrow AI, most Silicon Valley startups in the AI-enabled robotics industry choose to focus on specific vertical applications and collect proprietary data as a defensive advantage.Piece-picking by robots is the best example. As the technology improves, we will see warehouse piece-picking startups enter manufacturing, food, and agriculture markets. Moving away from manually programming a robot and having it learn and act autonomously on specific simple tasks — and sharing the learning — will become more important in the future.However, it’s not easy to find scalable and technically feasible business cases like piece picking that customers are willing to pay for. For example, assembly tasks vary from factory to factory, making it difficult to get sufficient training data or scale a solution across different customers.Given current technology constraints, there’s still a trade-off between precision and flexibility. Deep learning allows robots to handle variation, but these AI-enabled robots still cannot achieve the same precision and accuracy as traditional robots, at least not today.Some startups are providing modules or technology stacks for other players in the ecosystem to mitigate the complexity and scalability issues mentioned above. New areas where startups are likely to enter and disrupt include teleoperation modules, controllers that allow for safer human-robot collaboration, human-centered AI protocols, and sensors that can quantify human behavior.Eventually, we may see unified operating systems for AI-enabled robots similar to Windows or Android in the consumer electronics space, or standard user interfaces that will fundamentally transform the fragmented robotics industry and further accelerate its disruption.Standard modules and interfaces will bring down the cost to enter the market and shorten the time to market. As DRL and robotics technology improves, we will see a boom across industries including agriculture, food processing, home, and even surgical robotics.How technology will transform the industry landscapeLed by industry leaders like Amazon and Ocado, a British online grocery store, retailers see efficient and cost-effective supply chains as a key to success and are incentivized to automate their warehouses for lower cost, lower management complexity, and higher scalability. While Amazon keeps its Kiva mobile robots and automated warehouse solutions for in-house use only, Ocado launched its fulfillment platform to help other companies store, pick, and transport groceries. But most retailers still rely on material-handling companies and system integrators to build and maintain an automated warehouse.Incumbent material-handling companies tend to exploit the current technology, focusing on high-revenue-potential retrofit projects. This gives startups an opportunity to enter from the lower-end markets, where incumbents are less interested, and thus would not respond immediately. Dozens of startups offer scalable and affordable solutions that address customer needs for robots that can handle high-variability, low-precision tasks, such as piece picking in the warehouses.On the other hand, large robotics companies have long focused on developing high-speed and more-precise solutions to serve their main customers in the automotive and manufacturing industries. The pursuit of absolute precision is contrary to the process of deep learning, where results improve over time through trial and error.Some robotics companies saw an opportunity to enter new segments that require robots to handle high-variability, but lower-precision tasks, and began to build in-house AI teams and explore partnerships with startups developing deep neural network solutions for robots.There are currently more than 20 warehouse piece-picking robot startups worldwide, including Osaro and Righthand Robotics. Most of these startups (between seed round and Series C funding) are working on pilot projects with incumbent material-handling, retail, and robotics companies.Software and services will also capture more values, while premiums in hardware will decrease. With this power shift, as well as potential mergers and acquisitions, the landscape will change significantly over the next few years.The future of AI-enabled roboticsThoughts vary on the effect that AI and DRL will have on the robotics industry. One scenario is the robotics industry will consolidate because of the AI network effect, which says results improve as the amount of training data they’re given increases. This idea is that whoever owns the most proprietary data will win the competition.However, new researches in the areas of Transfer Learning, Meta Learning, and Zero-Shot Learning make it possible to learn from fewer and fewer examples, generalizing learning results across different fields. Cloud-based AI and robotics will also help accelerate learning and diffusion.As these technologies evolve, the landscape may face yet another significant change. This all starts from the warehouse!

The impact of artificial intelligence within warehouses will act as a crystal ball for AI, robotics in other industries

Why is Warehouse Automation So Important for AI Professionals?

Robots aren&rsquo;t yet household fixtures, but Georgia Tech researchers have already come up with a way domestic bots might recognize materials around the home.Using near-infrared light, similar to what&rsquo;s used in TV remotes, the robot can identify common materials used in household objects to better inform its actions. This might allow intelligent machines to understand, for example, the right bowl (paper versus metal) to put in a microwave or how hard to grasp a cup made of glass versus plastic.To classify materials, the researchers first determined hundreds of light wavelengths reflected from five common materials &ndash; paper, wood, plastic, metal, and fabric. With this information, they trained a neural network on 10,000 examples in order to create a machine-learning (ML) model that could be used by a robot to quickly identify a material.According to the researchers, a robot using their new ML model can identify materials without it first having to touch an object, a useful function for handling potentially fragile items. To do so, the robot holds a small spectrometer near an object to get a quick light measurement, which is then processed to identify the material.<iframe src="https://www.youtube.com/embed/fBv_xEai2AU?&wmode=opaque" allowfullscreen="" class="fr-draggable" style="width: 767px; height: 425px;" frameborder="0"></iframe>Classification of Household Materials via Spectroscopy&ldquo;Robots currently use conventional cameras or haptic sensing - the sense of touch - to estimate a material type,&rdquo; said Zackory Erickson, the first author on the research paper detailing the new work and Georgia Tech robotics Ph.D. student.&ldquo;This is the first time that we know of that spectroscopy and machine learning have been used for material classification in robotics research, and our accuracy is on par with existing methods.&rdquo;The team&rsquo;s new ML model yielded the best results using spectrometer measurements from near-infrared light. In fact, the accuracy was 99.9 percent with the full dataset of 10,000 measurements from 50 objects that the model had been trained on.&ldquo;While human eyes typically use three color receptors to see the world, our robot can be thought of as using hundreds of color receptors to recognize materials,&rdquo; said Charlie Kemp, associate professor in the Wallace H. Coulter Department of Biomedical Engineering at Georgia Tech and Emory University and part of the research team. &ldquo;Instead of a conventional color camera that measures red, green, and blue light, our robot uses a spectrometer that measures light at hundreds of different wavelengths, some outside of the range of human vision.&rdquo;To see how results would compare using only a single light reading from each object, the team also trained the model on just 50 measurements, one from each object. Interestingly, accuracy in identifying the correct material only dropped to 95 percent. When using a spectrometer reading from objects the machine learning model had never seen, the robot still achieved an 81.6 percent success rate.&ldquo;Spectroscopy presents a reliable and effective way for robots to estimate materials of household objects,&rdquo; Erickson said. &ldquo;We&rsquo;ve demonstrated how a robot can use near-infrared spectroscopy to infer the materials of everyday objects like cups, bowls, and garments.&rdquo;Reference:The research is published in the Proceedings of the 2019 International Conference on Robotics and Automation (ICRA) in the paper titled Classification of Household Materials via Spectroscopy co-authored by <a href="http://zackory.com/" target="_blank">Zackory Erickson</a>, Nathan Luskey, <a href="https://www.cc.gatech.edu/~chernova/" target="_blank">Sonia Chernova</a>, and <a href="http://charliekemp.com" target="_blank">Charlie Kemp</a>.

 Robot Classifies Materials of Household Objects Using 'Light-Reading' Device 

No contact is required with objects by using inexpensive, handheld 'light-reading' device

Robot able to instantly identify household materials using near-infrared light

The current record for a bicycle travelling across flat road is 133.78 km/h, set in 2012 by a Dutch team at the World Human Powered Speed Challenge, which takes place every year in the Nevada desert. But last September, a team from IUT Annecy aimed to beat that record. The team used artificial-intelligence-based software developed by <a href="https://www.neuralconcept.com/" rel="noopener noreferrer" target="_blank">Neural Concept</a>, to boost the performance of its bike. In just a few minutes, Neural Concept’s technology can calculate the optimal shape of a bike to make it as aerodynamic as possible. It can also be used for aerodynamics calculations in a number of other applications. The company is presenting its software in Stockholm today at the International Conference on Machine Learning.<iframe src="https://www.youtube.com/embed/9u6EyLiytJk?&amp;wmode=opaque" allowfullscreen="" class="fr-draggable" style="width: 771px; height: 437px;" frameborder="0"></iframe>From the outside, the IUT Annecy team’s recumbent bike looks more like a tiny racecar than a human-powered bicycle. It was custom-made to fit closely to the cyclist’s body. During the Challenge, he will have to ride down a 200-meter stretch of straight, flat road as fast as possible, after a run-up of 8 km. The design objective clearly isn’t cyclist comfort, but making the most out of every inch of the vehicle.Coming up with faster, more detailed and more effective designsExisting aerodynamic design methods require an enormous amount of computing power. Traditionally bicycle engineers think up different forms and then test them using computer simulation. But here, for the first time, engineers employed optimization software – rather than their own intuition – to define the recumbent bike fairing. The IUT Annecy team used Neural Concept’s software, specifying the bike’s maximum length and width and the space needed for the drivetrain and wheels. The program then sorted through all kinds of shapes, quickly comparing them in order to come up with the best one. For instance, the program helped the engineers determine the best location for the vehicle’s maximum width.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1562770461847-1920x1080.jpg" style="width: 766px;" class="fr-fic fr-dib">Pierre Baqué, CEO of Neural Concept © 2018 Alain Herzog&nbsp;To develop the technology behind the software, researchers at EPFL’s <a href="https://cvlab.epfl.ch/" rel="noopener noreferrer" target="_blank">Computer Vision Laboratory</a> trained a convolutional neural network to calculate the aerodynamic properties of various forms represented by generic polygon meshes, which are collections of points used to generate 3D shapes. This type of artificial intelligence works by running through several layers, categorizing information from the simplest to the most complex. In the initial layers, the program identifies a shape’s contours; then it assigns the contours to an object and determines what category the object belongs to based on the expected outcome.Engineers can use the software to carry out detailed analyses of different designs more rapidly and with better accuracy. “Our program results in designs that are sometimes 5–20% more aerodynamic than conventional methods. But even more importantly, it can be used in certain situations that conventional methods can’t,” says Pierre Baqué, CEO of Neural Concept. Another benefit is that the software can compare designs without human bias. “The shapes used in training the program can be very different from the standard shapes for a given object. That gives it a great deal of flexibility,” adds Baqué.The World Human Powered Speed Challenge is a competition involving bicycles designed by teams of university students. The Challenge is a real-world test for both the IUT Annecy team and Neural Concept’s machine-learning technology. The software has myriad other potential applications too, such as for designing drones, wind turbines and aircraft. Other industry professionals clearly see its potential – Baqué has been invited to speak at the world’s biggest machine learning conference today in Stockholm. IUT Annecy and Neural Concept have already started working on the bike for next year’s race. It will be designed exclusively and entirely by the software, without any human intervention.

Thanks to software developed bicycle engineers can quickly calculate the most aerodynamic shape for a bike. The software applies artificial intelligence to a set of user-defined specifications. Engineers have used the program to design a bike that they hope will break the world speed record.

Artificial intelligence helps design an ultra-aerodynamic bike

This is the first article, in a series, exploring the vast landscape of artificial muscle actuators. In the subsequent articles, the quest to create and harness artificial muscle technology will be explored.Until recently, conventional robots have utilized rigid materials and fabrication methods. Ironically, traditional robots are relatively weak and brittle, when compared with biological organisms, which utilize soft adaptive tissues to produce movement. A new approach to robotics has taken shape in the form of soft robotics and biomimicry.So, what does “soft” robotics mean? If you ever stayed up late as a child and remember watching old science fiction movies, as I did, your immediate concept of a “Robot” is typically a large, metal creature, lumbering about. In sharp contrast to this classic robot concept, soft robots are flexible and agile; think of a ballerina or an athlete. Obviously, modern robots have advanced from this stereotype, however our best robot is still not capable of the nuanced movements and physical strength of a human.A key component to producing a robot that more closely replicates the movement of a human, starts with the actuation system. An actuator is any mechanism that can convert some form of energy into physical movement, similar to how our muscles work. Typically, servo motors are used to produce movement in most robotic systems. However, servo motors are bulky, have a low power to weight ratio(when compared to muscle tissue) and are non-compliant. Essentially, compliance of a material means that it has some give to it, like a rubber band. You may not realize it, but the human body is capable of withstanding incredible forces.As demonstrated at the 2015 DARPA Robotics Challenge, a humanoid robot struggled to accomplish physical tasks that any average human would be able to do with ease. Common bumps and bruises that humans experience, would be catastrophic to most modern robots. Our ability to adapt to different physical environments is the template that will be used to create future robots.Artificial Muscle Research Engineers have sought a multitude of materials and fabrication methods in order to replicate the attributes of biological muscle tissue. I have also spent close to 20 years attempting to create an artificial muscle actuator, with limited success. So, what makes muscle tissue so special and replication of its attributes so elusive? Well, for starters, nature has had a 560 million year head start. Additionally, each of our muscles is a compact and powerful actuator that is made of soft, self-healing tissue. Our muscle tissue affords us dynamic movement and uses an efficient chemical fuel as a power source, glucose. Muscle tissue is composed of thousands of contractile fibers-within each fiber are up to 100,000 molecular units that synchronously shorten in length with a tremendous amount of force and velocity. This force and velocity enables basic movement and locomotion.Extensive research has been conducted to create an artificial muscle actuator which can replicate or exceed biological muscle tissue, with varying levels of success. Several types of technologies that have been explored include: Electroactive polymers (EAPs), Shape-memory metal alloy wires/liquid crystalline elastomers (SMAs), the Mckibben air muscle (Pneumatic artificial muscles), nylon fiber (torsional actuators) and Carbon fiber (PDMS rubber/torsional actuators). While all of these types of artificial muscle actuators, possess some of the sought after attributes of biological muscle, none of them are a complete solution in and of themselves. In order to fully appreciate the vast landscape of artificial muscle research and current innovations-each of these technologies will be further discussed; including history, fabrication techniques, advantages, drawbacks and future applications.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1561300296387-materiability2.jpg" style="width: 752px;" class="fr-fic fr-dib">EAP(Electroactive Polymer) ActuatorElectroactive polymers (EAPs) EAP technology is based upon the principle of generating a motive force in polymeric materials via an applied electrostatic field. One of the basic principles behind this technology is static electricity, if you have ever heard the crackle when you remove laundry from the dryer or accidentally, “shocked” your friend when walking, you have experienced static electricity. The electroactive effect occurs when a high voltage is applied to a flexible material, the electric static field that is generated is strong enough to produce movement in the material.Another technique used for actuation is based on elastomeric, meta-materials that exhibit the piezoelectric effect. Piezoelectric materials are actually very common, they are the technology that enables, haptic feedback in your cell phone. Typically, Piezoelectric materials are in the form of quartz crystals, which have the incredible ability to convert physical stress into electricity and vice versa. Moreover, certain polymer compounds can be, “doped”, with certain metals to produce the piezoelectric effect.&nbsp;While EAPs can embody, several permutations, the main focus of this discussion will be on the most commercially viable configurations which include, dielectric elastomers and ionic polymer-metal composites. EAP research started as early as the 1800s, however it wasn't until the turn of the century that EAP technology was realized for practical applications.Physicist, Dr. Yoseph Bar-Cohen, of the Jet Propulsion Laboratory, further advanced EAP technology by inspiring a global research initiative which challenged research groups around the world to compete in the world's first, “Arm wrestling Match of EAP Robotic Arm against Human challenge”. In 2002, Japanese tech company Eamex, released one of the first commercial products-a realistic robotic fish toy, utilizing a viable EAP, artificial muscle. Continued research, funded by DARPA, resulted in the tech start-up; Artificial Muscle, Inc-which specializes in development and manufacture of products based on EAP technology.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1561300403609-image022.jpg" style="width: 748px;" class="fr-fic fr-dib">Arm wrestling Match of EAP Robotic Arm against Human challenge(human opponent, Panna Felsen)Dielectric Elastomeric Actuator Dielectric elastomeric architecture involves an elastomeric film or cross sectional area of polymeric material, coated with an electrically conductive layer on either side. This configuration yields a compliant capacitor, where in, the conductive coating serves as a pair of electrodes. When an electrical field is applied to the DEA, changes in the electric field distribution of the electrodes generate a mechanical strain via Coulombic charge interaction of the compliant electrodes. Essentially, the conductive electrodes become electrostatically attracted and repelled in specific regions, causing the material to compress and expand. Additionally, the strain deformation state can be held in a static position via a constant DC voltage; analogous to an isometric muscle contraction.Composition of a DEA is relegated to a suitable elastomeric material with high elastic modulus; typically silicone or acrylic elastomers are used. During the fabrication process, elastomeric film of a predetermined thickness, is stretched upon a rigid frame in order to provide bidirectional static prestrain(stretch) of the elastomeric film. This prestrain fabrication technique has been shown to significantly increase the &nbsp;mechanical efficiency and responsiveness of DEAs. The electrically conductive coating, typically comprises conductive carbon grease or paste, applied to either side of the elastomeric film to provide diametrically opposed electrodes. This fabrication process can also be repeated several times, with each elastomeric film/electrode layer, “stacked” onto the subsequent layer in order to provide an actuator capable of exerting much larger forces.Ionic Polymer-Metal CompositesIPMC actuators are characterized by synthetic polymers engineered with positively charged metal ions, which are randomly arranged in an aqueous, polymer matrix. Typically, a noble metal coating is applied to the polymer-metal matrix in order to provide positive/negative electrodes(orientation of cathode/anode is based upon the terminal electrical wires attachment). If you remember from Chemistry class, noble elements are those that do not play well with other elements or resist chemical interaction. The positive metal ions partially entangle with water molecules and any inherent, negatively charged ions within the polymer matrix. When an electrical field is applied, the conjugated metal ions are attracted to the negatively charged metal electrode. This ionic migration of entangled water molecules provides an osmotic pressure shift resulting in mechanical strain across the polymer matrix membrane. Conversely, mechanically bending or deformation of the polymer matrix can produce electrical potential and output voltage, which can be advantageous for energy harvesting and biomedical sensors.Edge Over Other ActuatorsPossessing many biomimetic attributes has warranted continued research into EAPs development. Similar to biological muscle tissue, Dielectric Elastomers have an inherent ability to store large amounts of mechanical energy and contract with a large amount of usable force. Another appealing attribute is measured by how well an actuator can operate under continued stress and adaptability under varied conditions. A crucial element when considering building a humanoid robot that can work along side humans. With this in mind, DEAs have been shown to operate within standard room conditions, without the need for additional considerations. A crucial element for any biomimetic actuator is the ability to respond quickly when activation stimuli is applied, DEAs continue to excel with rapid response activation times measuring in the milliseconds.Additionally, Ionic Polymer-Metal Composites bring a unique set of properties to the table, including bi-directional actuation. Dependent upon the polarity of the activation voltage applied, IPMCs can reverse actuation direction, contributing further to adaptable mobility. Due to the unique chemistry of IPMCs, the mechanism of movement is attributed mostly to molecular motion, which allows for low activation voltages. An attribute that is extremely appealing in the quest to create a sustainable, autonomous robot. Configuration of IPMCs is not only constrained to simple membranes. There complaint nature allows the material to be rolled up into compact cylinders, with the ability to bend or extend. A quality that is important in order to approximate various muscle shapes and sizes, inherent in biological organisms.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1561300566769-a-Optical-image-of-the-bending-actuation-of-the-i-EAP-actuator-under-AC-electric.jpg" style="width: 300px;" class="fr-fic fr-dib">(a) Optical image of the bending actuation of the EAP actuator under AC electric signals.(b) Schematic of bending actuation mechanism.Challenges In The FieldWhile EAPs do posses several remarkable properties, we must also evaluate the engineering challenges faced in the development of EAPs for practical applications. One such hurdle to overcome for any robotic system is power consumption, especially within the context of realizing a self-contained humanoid robot. Unfortunately, this is a major con for DEAs, which require thousands of volts of electricity to operate. With the current state of battery technology this is far from ideal. EAP longevity is also limited due to the high voltage field, often exceeding the breakdown threshold of the polymer material. While IPMCs require minimal activation voltage, additional considerations are required for typical usage. Ironically, not unlike a biological muscle, IPMCs also require encapsulation in an impermeable membrane to prevent dehydration of the polymer matrix. In addition to IPMCs low actuation force, the most critical disadvantage remains the current inability to commercially, mass produce, a consistent ionic polymer. Until these challenges can be overcome, a viable EAP artificial muscle actuator remains elusive.Looking AheadWhat's in store for EAP technology? While an EAP based artificial muscle actuator seems unlikely in the immediate future. EAP technology is currently in development towards, biomechanical energy harvesting. Specifically, IPMC films can be attached to key flexing areas of the human body in order to harvest kinetic energy. So, next time you go to the gym, you can burn calories and charge your cell phone. This technology could also turn out to be invaluable for Soldiers in the midst of combat with little to no recharge options. The structure and mechanism of DEAs has found a practical application within the realm of virtual reality systems. An array of actuator “pixels” can be utilized for graphic displays for the visually impaired and possibly for theme park applications, such as virtual-haptic, hybrid environments. The National Science Foundation is continuing research towards an artificial muscle application, however. With the aid of 3-D printing technology, IPMC technology may prove to one day be a suitable artificial muscle solution.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1561300675742-3D_printed_IPMC_hand.jpg" style="width: 750px;" class="fr-fic fr-dib">A small 3-D printed ionic polymer-metal composite soft robotic hand. National Science Foundation researchers are working to transform this material into artificial muscles. This is the first article in a series&nbsp;about artificial muscles, created by&nbsp;<a href="https://www.wevolver.com/profile/samuel.manriquez" rel="noopener noreferrer" target="_blank">Samuel Manriquez</a>.

Skeleton with an artificial bicep. Credit: Aslan Miriyev - Columbia Engineering

The current state of robotic actuators and mechanical control systems are failing to keep pace with advancing AI software, exploring novel actuator technologies may bridge the gap between man and machine.

Biomechanical marvel-creating an artificial muscle actuator

Artificial intelligence (AI) is a wide-ranging tool that enables people to rethink how we integrate information, analyze data, and use the resulting insights to improve decision making


Artificial Intelligence: A Comprehensive Guide to its Engineering Principles and Applications

A.I.

CUDA Cores and Tensor Cores are specialized units within NVIDIA GPUs; the former are designed for a wide range of general GPU tasks, while the latter are specifically optimized to accelerate AI and deep learning through efficient matrix operations.

Tensor Cores vs CUDA Cores: The Powerhouses of GPU Computing from Nvidia

Automation and Industry 4.0 are revolutionizing the manufacturing and industrial sectors.

Everything you need to know about  various types of automation and their role in modern manufacturing systems. 

6 Types of Automation: A Comprehensive Guide for Engineers

In today’s digital age, where information is massively booming, ordinary softwares cannot handle the data because it’s Big Data. In this article, we’ll understand the main characteristics of Big Data i.e., the 3V’s, its technological advancements, applications, challenges, and data security.

Big Data: The 3 V's of Data

An AI-generated depiction of a modern day chipset

Gaining a comprehensive understanding of the distinct capabilities and applications of TPUs and GPUs is crucial for developers and researchers aiming to navigate the complex and rapidly changing terrain of artificial intelligence.

TPU vs GPU in AI: A Comprehensive Guide to Their Roles and Impact on Artificial Intelligence

This article provides an in-depth exploration of the evolution, fundamental concepts, diverse types, operating principles, and practical applications of ASICs.

What is an ASIC: A Comprehensive Guide to Understanding Application-Specific Integrated Circuits

Nvidia’s CUDA cores are specialized processing units within Nvidia graphics cards designed for handling complex parallel computations efficiently, making them pivotal in high-performance computing, gaming, and various graphics rendering applications.