Profile cover photo
Profile photo
Willy Rempel
Willy's posts

Post has shared content

Post has shared content

Post has shared content
I'll like to share a nice demo of a DL classifier running on a Nvidia Jetson TK1 for object recognition in (nearly) real time

Post has shared content
Dear all

Let me draw your attention to some work done by our group over the past few years in developing networks of neurons with

lumped dendritic nonlinearity and binary synapses. The connections on each dendritic branch is sparse and the neurons learn

to classify patterns when trained using different structural plasticity rules. Thus learning happens by modiying network

connection matrix and not by weight change. These networks have advantage for hardware implementations--reduced memory and

resistance to mismatch due to binary synapses.

We have demonstrated supervised and unsupervised spike time based learning rules and have used these networks as readout of

liquid state machines, to improve the reservoir in a liquid state machine, to classify spike latency patterns like

tempotron or to classify MNIST images. I summarize the papers below and also include arxiv links for your reading pleasure.

1. S. Hussain, S. C. Liu and A. Basu, "Biologically plausible, Hardware-friendly Structural Learning for Spike-based

pattern classification using a simple model of Active Dendrites,"   Neural Computation , vol. 27, no. 4, pp. 845-897, April


** Margin based learning, spike time based supervised learning, results on UCI dataset

2. S. Roy, A. Banerjee and A. Basu, "Liquid State Machine with Dendritically Enhanced Readout for Low-power, Neuromorphic

VLSI Implementations,"   IEEE Trans. on Biomedical Circuits & Systems , vol. 8, no. 5, pp. 681-695, 2014.


** Readout of LSM, 30X less binary synapses compared to parallel perceptron

3. S. Roy, P. P. San, S. Hussain, Lee Wang Wei and A. Basu, "Learning Spike Time Codes through Morphological Learning with

Binary Synapses,"  IEEE Trans. on Neural Networks & Learning Systems , vol. 27, no. 7, July 2016.


** Classifying spike time latency patterns, comparison with tempotron

4. S. Hussain and A. Basu, "Multi-class Classification by Adaptive Network of Dendritic Neurons with Binary Synapses using

Structural Plasticity,"  Frontiers in Neuroscience , Mar, 2016. doi: 10.3389/fnins.2016.00113


** Ensemble learning, adaptive allocation of dendrites, performance on MNIST, hardware savings analysis

5. S. Roy and A. Basu, "An Online Unsupervised Structural Plasticity Algorithm for Spiking Neural Networks,"  IEEE Trans.

on Neural Networks and Learning Systems , accepted, 2016.


** unsupervised spike time based learning, sequence learning

6.S. Roy and A. Basu, "An online structural plasticity rule for generating better reservoirs,"  Neural Computation ,

accepted, 2016.  


** Improving separation property of reservoirs in LSM


Arindam Basu
Assistant Professor
School of EEE
Nanyang Technological University

Post has shared content

Post has shared content
Tiny MEMS gravity sensor could detect drug tunnels, mineral deposits could trigger commercial revolution larger than MEMS motion sensors
A new device the size of a postage stamp can detect 1-part-per-billion changes in Earth’s gravitational field—equivalent to what the gizmo would experience if it were lifted a mere 3 millimeters. The technology may become so cheap and portable it could one ...

Post has shared content

Post has shared content
Biology software promises easier way to program living cells

Erika Check Hayden

"'Cello' automates the fast, reliable design of DNA-based logic circuits.

Synthetic biologists have created software that automates the design of DNA circuits for living cells.

The aim is to help people who are not skilled biologists to quickly design working biological systems, says synthetic biologist Christopher Voigt at the Massachusetts Institute of Technology in Cambridge, who led the work. “This is the first example where we’ve literally created a programming language for cells,” he says.

In the new software — called Cello — a user first specifies the kind of cell they are using and what they want it to do: for example, sense metabolic conditions in the gut and produce a drug in response. They type in commands to explain how these inputs and outputs should be logically connected, using a computing language called Verilog that electrical engineers have long relied on to design silicon circuits. Finally, Cello translates this information to design a DNA sequence that, when put into a cell, will execute the demands.

Voigt says his team is writing user interfaces that would allow biologists to write a single program and be returned different DNA sequences for different organisms. Anyone can access Cello through a Web-based interface, or by downloading its open-source code from the online repository GitHub.

”This paper solves the problem of the automated design, construction and testing of logic circuits in living cells,” says bioengineer Herbert Sauro at the University of Washington in Seattle, who was not involved in the study. The work is published in Science.1

Working together
Creating Cello required a decade of labour, says Voigt. The hard part, he says, was not writing the software itself but working out how to make biological parts — logic gates, by analogy with electronic circuits — that reliably worked together to carry out the functions programmed into the circuit by Verilog. For instance, the team had to develop a combination of genetic components that work together as an 'insulator' — ensuring that each biological part works no matter where in the DNA sequence it is placed.

How to design reliable, complex biological computing systems is a central problem of synthetic biology, Voigt says. Researchers found that DNA-based analogues of electronic switches and transistors would work in simple cases but would often fail when organized into more complex circuits. But inexpensive gene-synthesis technologies and the use of fast, cheap genetic sequencing to look at what goes wrong have enabled huge advances in understanding.

Living factories of the future
“What we’re finding over time is that biology isn’t this kind of mysterious unpredictable substrate; it just felt that way because we didn’t really have the tools to see what was going on,” Voigt says.

Voigt’s team tested 60 designs made using Cello; 45 worked correctly the first time. He estimates that it would take about a week to design 60 biological circuits with Cello; by contrast, it took a postdoc three years to design, test and build one successful biological circuit for a paper published in 2012, he says2.

Synthetic biologist Adam Arkin at the University of California, Berkeley, who was not involved in the work, says that Cello is one of a series of steps that aim to push synthetic biology closer to meeting its founding goals of using the principles of engineering to enable the design of new biological circuits.

“What is wonderful is seeing the original conception of synthetic biology — to build the infrastructure to make the engineering of new biological function vastly more efficient, predictable, transparent and safe — come to fruition in such powerful computational tools and biological reagents,” Arkin says..."

Post has shared content

Post has shared content
Tabor Manor founded by My People. Happy to see this
Brock Rec and Leisure students partner with Tabor Manor for intergenerational initiative
Wait while more posts are being loaded