Past study [5] has shown that retailers use conventional techniques with available data to model consumer purchase. While these help in estimating purchase pattern for loyal consumers and high selling items with reasonable accuracy, they do not perform well for the long tail. Since multiple parameters interact non-linearly to define consumer purchase pattern, traditional models are not sufficient to achieve high accuracy across thousands to millions of consumers

Was ist Byte-Ordering?

Alle konventionellen Rechner sind Byte-Adressiert. D.h. das Worte (egal ob 8, 16 oder mehr Bit) bestehen aus einer Folge (aufsteigender) Bytes. Dabei gilt das erste Byte als die Adresse des Wortes. Nimmt die Wertigkeit mit aufsteigender Adresse zu, ist es das Litte-Endian-Format, umgekehrt das Big-Endian-Format.

void* malloc(size_t size); // declared in stdlib.h

• Allocates memory block of size bytes
• Returns pointer to start of memory block
• Content of memory block undefined

Defining struct point and type Point for it

typedef struct point Point;
struct point {
    int x;
    int y;
};

CPU-memory bottleneck is also called von-Neumann-Bottleneck

Computer performance usually limited by

• Memory latency
• Memory bandwidth

➜ Memory bound vs. compute bound workload

Physical memory size affects latency because Signals have further to travel

Principle of locality A program usually accesses small portion of address space at any time (access pattern rather local than wide)

Temporal locality Location is accessed ➜ same location likely to be accessed soon again

Spatial locality Location is accessed ➜ nearby locations likely to be accessed soon again

Caches exploit temporal locality by remembering contents of recently accessed locations
Caches exploit spatial locality by fetching data around recently accessed locations

Tags and valid bits

How to know which memory block is stored in which cache block?

• Store block data and block address
• Only need high-order bits: tag

What if no data in a cache block?

• Valid bit: 1 = present, 0 = not present
• Initially 0

Direct mapped (1-way associative)
One possible location for placement

n-way set associative
n possible locations within a set

Fully associative (1 set)
Any location

Associativity	Location method	Number tag comparisons
Direct mapped	Index	1
n-way set associative	Search all entries within set (set given by index)	n
Fully associative	Search all entries	# entries

Block replacement strategies

Choice of entry to replace on a miss:

• Least recently used (LRU)
  • Complex and costly hardware for high associativity
• Pseudo-least recently used (PLRU)
• Random

Average memory access time = Hit time + Miss rate × Miss penalty

Compulsory miss (cold-start miss)

• Miss on very first access to a block which was never in cache before

Capacity miss

• Occurs when cache cannot contain all memory blocks accessed during program execution

How to reduce compulsory misses?

Solutions: Increase block size

Positive:

• Every block contains more data
  -> Better exploit spatial locality

Negative:

• Fewer but larger blocks
• Conflict misses might increase
• Capacity misses might increase

How to reduce capacity misses?

Solution: Increase cache size

Positive:

• more blocks in cache

Negatives:

• Hit time might increase (wire delays)
  -> L1 caches grow slowly, if at all

Conflict miss (collision miss)

• In direct-mapped or set-associative cache only
• Miss eliminated in fully-associative cache of same size as direct-mapped/set-associative cache

How to reduce conflict misses?

Solution: Increase associativity

Positive:

• Less collisions

Negatives:

• Critical path for access might become longer (selection of block)
  → Hit time might increase

Confusion matrix

y-axis -> true label

Confusion matrix
x-axis -> predicted label

Preprocessing steps (preparing data for neural networks):

1. Turn all data into numbers
2. Make sure your tensors are in the right shape
3. Scale features (normalize or standardize) Neural networks tend to prefer normalization.

Scale features (normalize or standardize)

Neural networks tend to prefer normalization.

Bag of tricks to improve model

1. Create model - more layers, more neurons, different activation

Bag of tricks to improve model

Compile model - other loss, other optimizer, change optimizer parameters