I hope this will be the first of a continuous, regular series of posts providing brief summaries concerning scientific literatures I recently encountered. By attempting to rephrase the encountered material I can probably better understand the underlying theory. It also allows me to keep a record of what I have done.
-
Quantum Mechanics for Scientist and Engineers
This is a book by David Miller. It is a very thorough encounter of QM often delving deeper into the mathemtaical formalisms then one would find in a regular physical chemistry book (e.g. Atkins). For example, it actually explains the origin of the Fermi’s Golden Rule and unpaired spin for fermions. There is also need concepts I have not met in my undergrad course such as the purturbation theory and transformation between the matrix and integral form of the representation.
I got a bit lost starting chapter 14 and stopped so I need to read on from there.
-
Could a neuroscientist understand a microprocessor?
It is a preprint paper at the time of writing. It begs the question that can fundamental insights be gained through more and better data? The investigators used a rather cool method of experimentation (in my opinion).
The method of investigation is via reverse-engineering of a simple processor unit (MOS6502), which has manageable number of transistors and then the processor unit’s behaviour was mimiced by a piece of simulational software. One can manipulate the processor to perform certain task which generates some output. One wishes to explain the functions of different components of the processor using the output. This is very similar to current available method to probe the brain.
The conclusion they drew was the inference about the processor functions was not always correct. From this, they concluded we would not necessarily construct correct models of the brain using our current methodologies. This reminds me of a recent article saying physics is not a pure experimental science, one always needs a hypothesis a priori in the precise form of mathemtics. Then the experiments can tell you whether the hypothesis seems to be correct. So for a system like the CPU or even more complex like the brain, would correlation based on large data be actually useful?
Although I guess using the processor as a way to test what methodology may not lead to correct conclusions for the brain due to the gross differences between the two systems.
-
A Neural Probabilistic Language Mode.
A feature vector representation was used for words. This way, all words are described by a set of descriptors each in real-number space. The set of descriptors is much smaller than the size of words. It quite dramatically outperformed trigram models in estimating the joint probability distribution of sentences. The trigram model could only account for context and only rather local contexts. The vector representation on the other hand could both account for words that are quite far apart and also words that are similar, both in terms of distance measures in vector space. However, the difference between these two situations cannot be differentiated purely from the distance, and it might be important to be able to do that. Essentially, learning the join distribution between word occurence has been transformed to learning the pairwise distance between each word descriptor vector.
The work uses a guessed vector representation for each word in the training set as input, which are transferred to a hidden layer (tanh activation) and then the output layer (softmax). This is a publication that guided the development of word2vec, marking it the first time neural nets in the context of NLP. However, I find it difficult to truly appreciate the issues, in particular speed penalties raised in this work and I think this is mainly due to my lack of practical experimentation with NLP methods.
-
Genetic Control of Biochemical Reactions in Neurospora
An early molecular genetics study using x-ray induced single-point mutations on neurospora. The idea was to associate mutation with observations of inability to carry out some essential metabolic processes (e.g. surcrose hydrolysis). This was observed by placing the fungi on different mediums lacking certain combinations of biochemicals that are known metabolites that the fungi can make. Hence, if a gene for encoding a certain process was mutated, then the fungi would not grow on medium deficient of that certain metabolite. It is worth pointing out genes have not been identified structurally and they were only a hypothetical entity at the time the paper was written. Therefore, it was only the authors’ assumption to think a gene is associated to one small step of a big biochemical cascade, rather than the entire process. In their words: “it should be possible, by finding a number of mutants unable to carry out a particular step in a given synthesis, to determine whether only one gene is ordinarily concerned with the immediate regulation of a given specific chemical reaction.” But such study could not elucidate the sequence of biochemical pathways.
-
Hydrogen Bond Basicity Prediction for Medicinal Chemistry Design.
The publication explores the empirical linear relationship between the measured hydrogen bond basicity (H-bond acceptor) and the electrostatic potential minima. By doing so, one can ultimately use hydrogen bonding complementarity for ligand design for binding to receptors. A very pragmatic paper, allowing the hydrogen bond acceptor values to be estimated when experimental measurements are not available, such as elaborate hetrocycles and peptides.
-
Data Science and Machine Learning with Apache Spark.
This is a three module lecture series offered by UC Berkely on edX, a relatively gentle lead into the area of data science using Apache Spark, which is a distributed framework to treat big data via parallelism. I participated in a similar offering about a year ago. Due to the dynamic and vibrant nature of the field, features have been amended and added since then. This time ,they have also chosen to have the assignment run on Databricks, who is the main contributor to Apache Spark, rather than having to configure a virtual machine to run the code last time which meant it is cumbersome to adapt the hardware architecture to perform customary jobs. Plus they are adding more content into the series.
To do this course requires some programming foundation in python. It is certainly very interesting to follow, though I suspect to fully appreciate the usefulness of the content requires at least a vague understanding of the time bottlenecks of manipulating massive datasets.