I think that all we need is an await tf.ready() statement in neuralNetwork init method. #105 (comment)

The ml5.setBackend function is an oddball because it initializes an asynchronous change but doesn't wait for it (tf.setBackend is async). So we need to wait for TF to be fully set up before using it.

Adding await tf.ready() seems to work! However, a new error occurred now that we are using webgpu backend:

Error: WebGPU readSync is only available for CPU-resident tensors.
    at ms.readSync (pose-detection.esm.js:17:87509)
    at Engine.readSync (engine.js:943:1)
    at Tensor.dataSync (tensor.js:297:1)
    at Tensor.arraySync (tensor.js:228:29)
    at NeuralNetworkData.js:509:1
    at engine.js:328:1
    at Engine.scopedRun (engine.js:338:1)
    at Engine.tidy (engine.js:327:1)
    at Module.tidy (globals.js:175:18)
    at NeuralNetworkData.createOneHotEncodings (NeuralNetworkData.js:494:12)

I will try to investigate this further.

Over the long term I think that the ideal thing would be to avoid synchronous operations and use .array() instead of .arraySync(). However this function is called way deep in a chain so there's a lot of methods that would need to be converted to async for that to work.

It seems like the error has been fixed by TFJS, not sure what version it was released in. It might be that we need to bump our dependencies.
Fix PR: tensorflow/tfjs#7576
Issue: tensorflow/tfjs#5468

Thank you @lindapaiste! I will try to bump the dependencies and see what happens. I agree that switching to async would be ideal over the long term.