44

This post has been updated a lot. On the top, you can see link updates. Below, variations on the initial answer. For the short version: successes of convolutional neural networks and deep learning look like a sort of Galilean revolution. For a practical point of view, classical signal processing or computer vision are dead... provided that you have enough ...


21

First, there is nothing wrong with doing grad work in image processing or computer vision and using deep learning. Deep learning is not killing image processing and computer vision, it is merely the current hot research topic in those fields. Second, deep learning is primarily used in object category recognition. But that is only one of many areas of ...


11

No Deep Learning isn't killing Image Processing. You need huge datasets and lots of computational resources to do deep learning. There are plenty of applications where it is desirable to be able to do image processing with less computational load and smaller memory footprints and without having access to huge databases. Some examples are mobile phones, ...


11

Today we had a discussion with a friend of mine. It was a rainy day here in Munich, while a large portion of Europe was having a kind of sunny atmosphere. People were sharing photographs in social media, where they were in nice summer dresses, wandering around the seas. She was annoyed with this situation and turned to me and asked: "Could you write a ...


7

The short answer is, No. DL can recognize a mug in a photo, but this doesn't kill signal processing in anyway. That said, your question is quite relevant in these troubled days. There is a nice panel discussion on the subject, featuring Stephane Mallat, etc., here.


6

Nvidia seems to have published some white papers comparing DNN inference performance between high-powered CPUs and (of course) Nvidia GPUs. (one example) Ballpark seems to be that some systems can meet or exceed typical video frame rate thru-put for some class of DNN classification tasks. Whether those image sizes and/or DNN architectures and ...


5

From feature extraction to learning the desired result, deep learning algorithms can act as full pipelines for solving tasks at hand. End-to-end learning usually refers to omitting any hand-crafted intermediary algorithms and directly learning the solution of a given problem from the sampled dataset. This could involve concatenation of different networks ...


5

Data engineering is still used in machine learning to preprocess and select the data fed to DNNs to improve their learning time and their evaluation efficiency. Image processing (the stuff between the camera sensor and the RGB/etc. bitmaps fed to the DNNs), a form of data engineering, is still needed.


4

I really don't do much image processing but I worked for an organization (US Navy) that did and funded research in signal classification the last time Neural Nets were a hot topic, the mid to late 80's. I had to sit through a large number of essentially marketing stuff. The arguments were along the lines of: It's Neural, like your brain and since it ...


4

A thorough understanding of signal processing (along with linear algebra, vector calculus, mathematical statistics etc.) is imo indispensable for non-trivial work in the field of deep learning, especially in computer vision. Some of the high impact papers in deep learning (now that most of the low hanging fruit have been picked) evince a good understanding ...


4

The power of complex representations remains an open topic to me. I still do strive the understand Fourier transformations. An underlying question is, to me: why would complex transformations be useful for real data? More generally, when data dwell in a set $S$, is $S$ the most appropriate set of analysis, or is it more appropriate to resort to a bigger ...


3

The kernels used by a ConvNet are nothing but neural weights. You can think of them as a multilayer perceptron with some connections cut off and some weights restricted to be equal (weight sharing). With this in mind, we must take into account that the kernels (or filters, in this context) are learned, so they depend exclusively on the type of inputs and ...


3

Training is essentially an optimization and assuming you will have complex weights $z$, and a real valued objective, $$ \min_{z} f(z) \quad f(z) \quad \text{Real} $$ because of the Cauchy Riemann Condition, $f(z)$ is not analytic (essentially doesn't have a Taylor Series in $z$) so this is handled as: $$ \min_{z,z^*} f(z,z^*) $$ where $z$ and $z^*$ are ...


3

My perspective from university was that many signal processing people were a bit hostile toward ML, I suspect because they felt threatened that it was encroaching on their domain. But recently there's been a lot of research into benefits of complex valued deep neural networks, which may suggest that the golden ticket is really a solid understanding of both ...


2

Well, yes. In the same way that the development in higher level programming languages like C++ and Python 'killed' assembly programming. That does not mean it is irrelevant to learn assembly when you enroll in a CS course however. It provides great insight in how the computer works, what goes on behind the scenes of higher level languages, what the basic ...


2

Here is a newly published paper and video example: https://www.youtube.com/watch?v=w2iV8gt5cd4 http://arxiv.org/abs/1411.4389


2

Since next layer is fully connected it does not really matter what shape your pooling output would be. You have 14x100, you can rearrange them as 1x1400 as input for next layer, 1000 elements as output. The paper says that they select the size of the fully-connected layer of 1000 to be close to the output size of the pooling layer.


2

There are as many ad hoc "definitions" of "real-time" are there are people or places you hear or read about what it supposedly means. Some people claim there is a "hardline" such as 10uS response time, but I believe there is no academic paper that makes that mistake. Many people implicitly have an informal mental model that considers a system as being "...


2

What you are trying to archive is very hard if at all possible. The pinnacle of research today (including my own) is striving for good results on such tasks. DCASE challanges engage with similar tasks with some interesting results, though the make assumption which cannot be generalized to your case. In DCASE2019 task 3, for example, they assume up to 2 ...


1

Your question can be interpreted in many ways. First, convolution with a number is not a common operation. Convolution is defined by: $$(f*g)=\int_{-\infty}^{\infty}f(\tau)g(t-\tau)d\tau$$ Assuming a constant signal $g(t)=a\in \mathbb{R}$ this is reduced to: $$(f*g)=\int_{-\infty}^{\infty}af(\tau)d\tau=a\int_{-\infty}^{\infty}f(\tau)d\tau$$ I am assuming,...


1

Are there any languages and vocabularies in both data sets? If so, examine the differences between data sets using only those samples common to both data sets. There could be differences in dialects, microphone response, equalization, background noise, room acoustics, HRTF, and etc.


1

First of all, yes, what can be achieved by any machine learning tool is only as good as the dataset it was trained on. The ML can't guess what properties of the high-dimensional data you want it to respect and which not. So, you'll have to make sure all your data sources are either equal in their probability distributions when it comes to the feature you'...


1

We can start from what is "shift invarient": Transform G is shift invariant if - $$\forall x:\sigma^nG(x) = G(x)$$ $\sigma^n$ being shift by n. Examples for transforms that are invarient to shifts are histogram and the amplitude of Fourier transform. Commuting with shift is - $$\forall x:\sigma^nG(x) = G(\sigma^nx)$$ So it can't be shift invariant (unless ...


1

"Real Time" is defined with respect to the problem. Typically, "Real Time" is any system that carries out any computation required to provide an output in time less than the sampling period $Ts$. To define "Real Time" for your application, you need a quantification of "time" or "flow" or "rate" so that you can compare that to what your system achieves or ...


1

One of the popular feature descriptors used for human detection is HOG - Histogram of Oriented Gradients. Usually you would train a classifier for recognizing human vs non-human, and then you would implement a sliding window technique. HOG has implementations in most of the scientific libraries in the Python community, for example in skimage there is an ...


1

In short, yes. One thing that you'll find is that the human ear is actually relatively insensitive to phase, so you can take an audio sample, apply an allpass filter that doesn't affect the signal's magnitude spectrum but distorts its phase, and it won't make it sound much different to you. However, the time-domain waveform could look completely different. ...


1

It is very huge effort to build RNN-HMM system with HTK. It is easier to use more flexible toolkits like Kaldi, they have RNN-HMM already implemented. HTK in latest version 3.5 beta already has support for DNN (not RNN but still a neural network), you can grep for HYBRIDHS in sources. For details on DNN-HMM it is better to read HTKBook 3.5 section 3.12 DNN-...


1

CNN Let's say you have $P$ phones $p_0,p_1,\ldots,p_{P-1}$ and the CNN generates the posterior probabilities $x_i = x(p_i)$ for $i=0,\ldots,P-1$. HMM A Hidden Markov Model is a system that has $S$ states $s_0,\ldots,s_{S-1}$ that can produce $O$ observations $o_0,\ldots,o_{O-1}$. The probability of jumping from state $s_i$ to state $s_j$ is $a_{ij}$, ...


1

I think what you would like to do is something similar to an Auto-Encoder, in which one would append an Encoder to reconstruct back the image. The 'kernel', or more appropriately the intermediary representation is used to synthesize an image, and an appropriate reconstruction objective is minimized to recover the input. On the other hand, you could always ...


1

Well, yes and no. There are multiple factors, such as the level of annotation and the accuracy. If one requires human-level annotation performance on completely unknown scenes (with no prior information about context), and wants to annotate the images fully autonomously, without human intervention, then probably yes. Image annotation generally involves ...


Only top voted, non community-wiki answers of a minimum length are eligible