Saturday July 2, 2022, 9:00-13:00 Mountain Time (GMT-5)

Organizers: Cornelia Fermuller, Terry Stewart, Tobi Delbruck, Shih-Chii Liu, Emre Neftci, Andreas Andreou

The panelists have been invited to prepare to respond to one or more of the questions below about the relation of neuromorphic engineering to technology and biology.

The forum will take place physically at the schoolhouse (with remote participation of some panelists) and will be livestreamed.

To participate online use this zoom link (http://tiny.cc/neuromorphzoom), and register at: https://sites.google.com/view/telluride-2022/online-registration?authuser=0

Questions related to Technology

I. Which approaches to neuromorphic electronic engineering were successful and what should we change?

Carver Mead said “Listen to the silicon, it tells you what it wants to do.” Based on this message, neuromorphic electronic engineering for decades focused on one interpretation of this message: Exploiting the similarity between subthreshold FET operation and exponentials in voltage sensitive channels in mixed signal circuits that do analog computation.

1. Has this approach paid off? If so, how? If not, why?

2. Was the message misinterpreted, or should we reconsider what it means given 30 years of experience?

3. What is the message we should now be hearing?

II. What are the next challenges for neuromorphic engineering systems?

E.g. What are your top 3 from the following list? Are there missing challenges?

1. Improving inference power, area, and throughput

2. Improving offline training of DNNs

3. Improving ability to build systems that continuously adapt and learn, e.g. by self-supervised learning

4. Developing new architectures for computing technology

5. Making neuromorphic computing platforms that can have an impact

6. Improving neuromorphic sensors and systems so that they are adopted by major player for mass production

7. Bringing in new memory technologies to PIM or CIM

8. Neuromorphics and hyperdimensional computing: A new approach for cognition?

III. Which approaches in mainstream AI have the potential for improvement via neuromorphic organizing principles?

E.g., rank the following approaches/architectures:

1. CNNs

2. RNNs

3. Convolutional recurrent networks

4. Attention, transformers, other recent mainstream AI architectures

5. Self-supervised (continuous?) learning

6. Hyperdimensional computing

7. Compute/Process in Memory (CIM/PIM)

8. Something missing from this list

IV. What real-world open problems can NE help solve for society?

1. Can neuromorphic engineering contribute to robotics, security, IoT, wearables, drones, prosthetics, combinatorial optimization problems, higher-level cognition?

2. Neuromorphic Engineering and Cognition: Does it matter at this time for higher-level processes whether we compute neuromorphic? Which processes would demonstrate the benefit of neuromorphic hardware computing?

3. Can spike-based computing have real industrial impact? Or is it already, via the exploitation of sparsity in many mainstream digital AI hardware?

4. What new hardware principles are expected to change neuromorphic intelligence?

5. Should we think about systems engineering for neuromorphic engineering design now?

Questions related to Biology

I. Neuromorphic engineering was posed by Misha Mahowald as a way to better understand biological neural computation. How can neuromorphic engineering better interact with neuroscience/cognitive science? (in terms of bi-directional interactions)

1. Has this approach to neuroscience paid off? If so, how? If not, why?

2. As a practicing neuroscientist, can you name organizing principles or ideas stemming from neuromorphic engineering and AI that have inspired or guided your experimental neuroscience?

3. How can neuromorphic engineers better work with neuroscientists and cognitive scientists?

II. Comment on successful or promising approaches for the future on the interplay of NE and studying/understanding intelligence:

1. How can neuromorphic engineering contribute to understanding intelligence?

2. What can neuroscience contribute to understanding intelligence?

3. At what level can/ should we distill organizing principles of biology and neuroscience to implement them in our software/hardware?

4. What are we missing in our models of cortical computation in our AI solutions?

III. Can neuroscience contribute to developing different and better computations (algorithms and architectures)?

1. Which “organizing principles” from the brain have not been exploited by neuromorphic engineers?

2. Which biological neural architectures should neuromorphic engineering reconsider investigating for their utility in artificial intelligence? E.g. cerebellum, motor cortex, amygdala, spinal cord?

3. Which cognitive tasks / principles should we study and seek to implement, for example for better inference, training, continuous adaptation and continual/incremental learning, and new architectures for computing technology?

4. How can we bring evolutionary adaptation, development and self-organization to technology and NE in a practical form, suitable for future mass production?