Dan Goodman

@neural-reckoning.org

3815 Followers 389 Following

Computational neuroscientist at Imperial College. I like spikes and making science better (Neuromatch, Brian spiking neural network simulator, SNUFA annual workshop on spiking neurons). 🧪 https://neural-reckoning.org/ 📷 https://adobe.ly/3On5B29

Long exposure photo of a little rocky island in a wavey sea that makes the water look like smoke enveloping the bottom of the island. In the distance background is a hazy mountain.
A long exposure of waves crashing into rocks. The long exposure makes the waves look like white smoke settling in pools around the rocks.
Black and white photo of the full moon with a few light clouds reflecting on the sea. High contrast so all you can see is the moon, reflections, lights of a distant town, and traces of some clouds.
Seagull silhouette in front of setting sun. The sun is behind some clouds and so appears as a distorted white blob in the center with yellow/gold streaks at the edges.
Hills over the sea just after dusk. Long exposure shot makes the sky and the sea a reflection of each other as if the water was still. The land is almost completely black with the just set sun creating a halo around 4 trees.
A bee flying between two branches laden with bright yellow flowers.
Wetland pond with dramatic sky fading from white to blue with dark clouds with bright white surrounds from the sun behind. There are some tall apartment blocks looking over the water, and reeds in front of the water.
Tree in bloom. White flowers glow luminously against a dark grey cloud background.
Rainbow peeking through behind trees lit up golden in sunset light, dark clouds behind.
Woodland with rich brown floor covered in leaves and sticks, and sun shining bright green through foliage above. On the floor there are trails of smoothed out mud lined with sticks forming paths.
Woodpecker on the side of a tree with blue sky behind and some out of focus green leaves in front.
Crow on a branch with a twig in its beak.
A fox on a pedestrian walkway at twilight. There are trees and bushes around in greens and yellows. One of the bushes has no leaves and is hung with silver baubles. The fox is sitting quietly, observing, between two walls.
Two feet at the bottom of the picture facing a sea of very sticky mud.
Notification saying "Hari Seldon followed you" 42m ago
Black and white photo of a winter tree with many crows in it that stand out in stark black contrast.
Silhouette of a sunset sky with a man looking up thoughtfully. There are trees and a variety of clouds ribboned across the sky, with orange yellow layers around the man, going into shades of grey, blue and white in the sky.
London skyline with dramatic lighting, buildings emerging from a layer of mist near the ground, orange and dark grey clouds above.
Full (or close enough) moon over a zigzagging park path with quirkily angled street lights and the orange flow of the setting sun in the trees, dark shadows below.
Dark blue sky, bright moon, pink contrail, black silhouette of tree.
Sunset over parliament hill, with a bright orange sun, blue sky with a handful of clouds with orange glow and dark grey interior casting shadows upwards. A few figures are silhouetted against the setting sun.
Box showing "Reading Assistant" With the following text: Questions you could ask: Why is it important to differentiate between artificial achievement and intelligence in Artificial Intelligence (AI) systems? What distinguishes the evolution of human intelligence from that of artificial intelligence? Why is artificial intelligence considered a computational construct rather than a psychological one? Actions you could take: Summarize this article Sign in to unlock the full response.
Two posts from Bluesky. The first one shows a figure from a paper published in Nature Scientific Reports full of totally incoherent AI fabricated gibberish words. The other a comment on a recently published paper by eLife discussing the paper and its peer reviews which were published along with the paper.
Late autumn photo of a tree by a pond. The leaves are golden brown on the tree and in a carpet underneath, a clear blue sky in the background.
Dashed line indicating how fast performance falls when you prune the longest delays, versus solid line indicating the same for shortest delays. Performance drops much faster when pruning long versus short delays.
Plot showing how accuracy at SHD doesn't depend much on number of bits used for weights as long as you have at least 3 weight values, and use at least 3 bits per delay.
Graph showing performance using just weights versus weights+delays as a function of total number of parameters for datasets with temporal complexity varying from low (DVS gestures) to high (NTIDIGITS, SSC, SHD). In all the high temporal complexity cases you get much higher performance (accuracy up to 30% higher) with delays than with weights, for the same number of paramters.
Leaderboard showing that 5 of the top 10 models for the SHD dataset have the word 'delay' in their title.
A hedgehog scientist cartoon saying "SNUFA will be starting soon"
Meeting chat showing everyone posting hedgehog emoji after I suggested as the official emoji of the workshop.
Screenshot of a website with spinning dots saying "Loading data... (loading speed will be improved next year)"
Halloween decorated house. Close up of red lit skeletal man with glowing yellow eyes, holding a lollipop in one hand held out, and a hammer in the other behind his back. In the background are other halloweed themed things and children in fancy dress. On the left, a parent is taking a photo with a phone.
A carved pumpkin lit from the inside with one tooth, and golf tees sticking out of its head.
Happy looking carved pumpkin with a big nose, a bit like a turtle. It has a little hat with a bat and a heart.
Photograph of an autumnal tree with sun behind. The trunk and branches of the tree are almost black, and the leaves glow yellow in the sun shining through them.
Trees with leaves of every shade from dark to light green, yellow, orange, brown, red and purple. Background is flat white overcast sky.
Squirrel on a concrete wall with patches of moss, a blurry street in the background.
Swan in water with bright sparkles of light behind as the sun reflects off the water.
Autumnal tree showing a gradient of every colour from dark to light green to yellow to orange to red to purple, with a blue sky in the background.
Black and white photo of an old mine cart against a lake district backdrop. The mine cart is rusting and falling apart and it's on a tiny section of track kept just to display it. In the background are some hills and a cloudy sky.
A histogram showing the time a decision was made with and without neuromodulation. You can see that without neuromodulation there is a unimodal peak around 700-800ms. With neuromodulation, the peak splits into two modes at around 400ms and 1000ms, and the mean drops by around 100ms.
Four panels. Top left panel shows that when there is no noise, average firing rates start low, increase a bit up to around 200ms in to the stimulus, then decrease slowly until the end of the stimulus at around 1400ms. With noise, and no neuromodulation, you can see that it already somewhat uses a listening in the dips strategy, with firing rates slightly higher at phases where the noise amplitude is low. With neuromodulation, this effect is very strong, and this holds at different amplitude modulation frequencies.
Graph showing speech in noise performance for various signal to noise ratios from -17 dB to quiet. Both with and without neuromodulation performs well at -5 dB (comparable to quiet), but performance starts to drop below that. By -17 dB the network without neuromodulation is at 35% accuracy (chance level is 5%) while the one with neuromodulation is at just under 60%.
Graph showing performance for different spatial grouping factors from 1 to 200. Performance is about the same for all values except the largest G=200 where it's a little bit less (but even then, not much less).
Three panels for the SHD, SSC and DVS tasks. Each panel shows a baseline performance without modulation, and three coloured lines for three different ways of doing the neuromodulation. There are slight differences between these curves but they're mostly the same. In each case, for most temporal scales performance is much better (SHD goes from 80% to 92%, SSC from 60% to 73%, and DVS from 78% to 85%). Peak performance is just below 100 ms time scale for SHD, at the smallest possible 1ms timescale for SSC and around 150 ms for DVS.
Temporal grouping diagram. Modulation effects are transmitted to target neurons every K steps of the simulation.
Architecture diagram. Input connects to a recurrent primary SNN layer and to a modulator network. The modulator network also receives input from the primary SNN, and its output modifies the parameters of the primary SNN. The primary SNN then feeds to the output layer.