“1.2.4 A SPECIFIC EXAMPLE OF TMBI’S INTERDISCIPLINARY RESEARCH
With over 270 scientists with permanent positions enrolled already in the TMBI project, it would totally impossible to present all the potential projects that we hope to promote over the next decade if the project has the status of a Convergence Institute. Some more detailed information can be found in some of the support letters at the end of this document.
However, to illustrate the sort of innovative research that we hope to promote at the TMBI, we will have a brief look at one particularly promising line of research that we are currently working on, and which is related to Thorpe’s ERC Advanced Grant project M4: Memory Mechanisms in Man and Machine. As mentioned earlier, there is enormous interest in the brain inspired computing, and in particular deep learning. This approach, which has received massive funding from major companies like Google and Facebook uses relatively simple feedforward convolutional neural networks to solve challenging problems that include object recognition in natural scenes. The main ideas, including the error backpropagation learning algorithm have been around since they were originally developed by people like Geoff Hinton and Yann Lecun in the mid 80s. However, with recent developments in computer hardware, it has now become possible to use billions of cycles of training with labelled data. In 2012, Hinton and his students completely outclassed the then state of the art in computer vision with a system trained over a few weeks. Since then, the approach has taken over – and Hinton is now working for Google. The problem is that, although the systems work very well, and can potentially replace human operators on many tasks, the way they learn is totally unbiological. Human children don’t need to be told millions of times “this is a dog” and “this is a cat” to learn to categorize. Thorpe and his team think that they have come up with a biological plausible learning scheme that learns like humans do – simply on the basis of repetition. They have an algorithm using a modified learning rule based on Spike-Time Dependent Plasticity that they have implemented on a chip called an FPGA (Field Programmable Gate Array) that costs just $100 but will learn arbitrary sensory patterns in just a few repetitions – just like humans. The technique has already been applied to visual and auditory stimuli which can be converted into spatiotemporal patterns of spikes using something similar to the human retina or cochlear. But the possibilities are endless. For example, we are now looking into the possibility of collaborating with researchers at the LAAS to use similar ideas in robotics. The industrial applications could be enormous, and major companies including Intel and Samsung are clearly interested.
At the beginning of September 2016, SpikeNet Technology, the company that Thorpe created in 1999 with Arnaud Delorme and Rufin VanRullen was acquired by Californian high tech start-up BrainChip Inc with the aim of producing hardware designs based on SpikeNet's bio-inspired learning algorithms. Very recently, BrainChip is in the process of signing a licensing deal with Toulouse Tech Transfer to obtain the exclusive right to commercialize products based on the new algorithms developed at the Cerco by Thorpe and his team. Given the enormous commercial potential of this sort of technology, the revenue sharing scheme could have a very major impact not only on the CerCo, but potentially also across the TMBI.
While this is only a very brief illustration, it demonstrates how an interdisciplinary approach that brings together researchers working both on biological and artificial systems can be so rewarding. For more illustrations of the sorts of innovative projects that TMBI hopes to foster, please refer to the “Projects” section of the TMBI website which will be updated in the coming weeks and months with yet more ideas.”
“The CerCo and CIC have strong ties to industrial partners like SpikeNET and Cochlear, that develop brain-inspired image processing systems and sensory aids. But TMBI will encourage even stronger links with AI & Computer Science. With the advent of deep learning, such collaborations will be increasingly bidirectional: not only can biology help provide engineering solutions for computer vision, artificial neural networks provide a reference model—and source of inspiration—for neuroscientific and psychological investigations”
“The council meets monthly, but roughly twice a year it will be supplemented by representatives of the CNRS, INSERM, the Federal University, Toulouse City Council, the Region, the Hospital and two representatives from industry, (for example, Airbus & BrainChip) bringing the total to 22.”
BRN Price at posting:
23.0¢ Sentiment: Buy Disclosure: Held
(20min delay)