We generated ‘model metamers’ – signals that produce nearly identical responses at some stage of a neural network. They thus produce nearly identical responses at all downstream stages, and the same decision. (2/n)pic.twitter.com/sDXr06n6wr
U tweetove putem weba ili aplikacija drugih proizvođača možete dodati podatke o lokaciji, kao što su grad ili točna lokacija. Povijest lokacija tweetova uvijek možete izbrisati. Saznajte više
We generated ‘model metamers’ – signals that produce nearly identical responses at some stage of a neural network. They thus produce nearly identical responses at all downstream stages, and the same decision. (2/n)pic.twitter.com/sDXr06n6wr
Model metamers generated from deep layers of a neural network capture the invariances of the network representations – physically distinct signals are mapped to the same downstream representation, as one might imagine would aid recognition. (3/n)
The logic is that if we have a good model of human perception, say of speech recognition, then if we pick two sounds that the model judges to be the same, a human listener should also judge them to be the same. (4/n)
The main result of the paper is that across both image- and audio-trained neural networks, model metamers generated from deep network layers are unnatural and unrecognizable to human observers. Here are examples. (5/n)pic.twitter.com/hsmRETMeHN
We quantified this with behavioral experiments, confirming that as one moves deeper into a trained network, model metamers become progressively less and less recognizable to humans (despite being perfectly recognizable to the model). (6/n)pic.twitter.com/yAiqSqJW74
The phenomenon is reminiscent of ‘adversarial’ stimuli – judged differently by a model but indistinguishable to humans – but seems like a more problematic inconsistency, because metamers are not adversarial. They are generic samples of what looks/sounds the same to a model. (7/n)
This synthesis method has been previously used to visualize neural network representations, but is often combined with priors to make the resulting images more naturalistic, masking the extent of divergence with humans. (8/n)
For me, these results point out the exciting challenges that await as we try to build complete models of human perception. Neural networks remain the best available models of human vision and audition by many metrics, but we have a ways to go. (9/n)
To this end, the results also suggest directions for making models more human-like. For instance, reducing the aliasing that is common in many neural network architectures makes model metamers more recognizable to humans. (10/n)pic.twitter.com/9rF8hLbW4Q
Lots more results in the paper, to be presented by Jenelle this Wednesday at NeurIPS. Full poster, models, code etc. available here: https://github.com/jenellefeather/model_metamers … Video summary here: https://www.youtube.com/watch?v=nFtHc0mxRhs … (end)
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.