1. Python is free, just like science. 2. Python is opens source. 3. Python is where programming is headed, everyone should use it. 4. People are lazy and just don’t want to port legacy code. 5. Python does deep learning though, which is hot. ?. Why do you hate Python, you hater?.

Reasons for switching covered in the post:.

Generate videos in just a few seconds. Try Grok Imagine, free for a limited time.

To whatever #Neuroscience lab needs to hear it: don't be pressured to switch from #matlab to python if it's not the right tool. You do you.

An argument for embracing the coming wave of neurotechnology: Neural augmentation can enhance and amplify our humanity without threatening our identity. After all, it worked for the Borg.

The output of Softmax should be interpreted as "the amount of class-N-iness" the input has, not the input's "probability of belonging to class N".

RT @FanChaofei: 4 yrs ago, I jumped from AI into Brain-Computer Interface (BCI) research—with almost zero neuroscience experience. Today,….

Thanks to Pedro Alcolea, @LeeEugene57, @Sean_LakeShore, Kevin Bodkin for making this happen.

If you do #BCI work, I would love to know if this decoder works as well for you as it did for us.

If you want more than a tweet but less than the paper, check out the blog post.

Discrete velocity decoding worked very well for a monkey using an intracortical #BCI as well. (37% target hits with the Wiener filter vs 61% with DDS).

The discrete direction decoder (DDS, purple) makes staircase-looking cursor paths, but people do really well with it in our iBCI model, hitting way more targets than standard decoders. tl;dr on the #BCI model we used (the jaBCI):

#BCI world, we made a new decoder that works *really* well. It picks a velocity off a short menu instead of allowing any possible 2D vel. 93% vs 39% success in our model and 61% vs 37% in NHP iBCI for cursor control. Does it work for your #BCI?. Paper:

Suppose you wanted to use shrewd monetary policy to responsibly make biomed research more efficient, what would you do?.For inflation, the FedReserve would deliberate for months and change int. rates by 0.25%. Here, without *any* warning, NIH changes rates by ~35% over a weekend.

Whatever you think about cutting overhead, slashing it this severely, unexpectedly, and suddenly will do more harm to biomedical research than it will do good for reigning in administrative bloat.

RT @CarlosdelRio7: Impact of NIH funding in Georgia: .NIH AWARDS FUNDING: $780 M.JOBS SUPPORTED: 11,816.ECONOMIC ACTIVITY SUPPORTED: $2.18….

For context, I have only been faculty at R1 universities and my work is in biomedical engineering and neuroscience.

Many science SM posts are about new grants and papers, which makes it hard to get a realistic sense picture of what its like applying for grants. To help, especially for #newPI, I'm sharing a record of my grant submissions. Details on the blog:

The secret sauce for this approach is to use sequential (Newton) optimization that at each step 1st solves for the system dynamics then 2nd updating the ANN parameters using the step-1 results. It works great in many simulated ODE systems.

It uses Kalman filters to train ANNs that are embedded inside systems of ODEs describing actual physics. The beauty is the ANNs get constrained by the ODE physics and learn to generate physically meaningful meaningful approximations.

Ever had an incomplete (ODE) model of a system but can't complete it because your system has some lurking states you can't record? We have a method to accurately infer the missing state *and* build a neural ODE for its equation! . It's at AAAI 2025: