#include "bits/stdc++.h"
#include <random>
#include <chrono>
#pragma GCC target("avx2")
#pragma GCC optimize("O3")
#pragma GCC optimize("unroll-loops")
#define ALL(x) (x).begin(), (x).end()
#define RALL(x) (x).rbegin(), (x).rend()
#define SZ(x) ((lint)(x).size())
#define FOR(i, begin, end) for(lint i=(begin),i##_end_=(end);i<i##_end_;++i)
#define IFOR(i, begin, end) for(lint i=(end)-1,i##_begin_=(begin);i>=i##_begin_;--i)
#define REP(i, n) FOR(i,0,n)
#define IREP(i, n) IFOR(i,0,n)
#define endk '\n'
using namespace std; typedef unsigned long long _ulong; typedef long long int lint; typedef long double ld; typedef pair<lint, lint> plint; typedef pair<ld, ld> pld;
struct fast_ios { fast_ios() { cin.tie(nullptr), ios::sync_with_stdio(false), cout << fixed << setprecision(10); }; } fast_ios_;
template<class T> auto add = [](T a, T b) -> T { return a + b; };
template<class T> auto mul = [](T a, T b) -> T { return a * b; };
template<class T> auto f_max = [](T a, T b) -> T { return max(a, b); };
template<class T> auto f_min = [](T a, T b) -> T { return min(a, b); };
template<class T> using V = vector<T>;
using Vl = V<lint>; using VVl = V<Vl>; using VVVl = V<V<Vl>>;
template< typename T > ostream& operator<<(ostream& os, const vector< T >& v) {
    for (int i = 0; i < (int)v.size(); i++) os << v[i] << (i + 1 != v.size() ? " " : "");
    return os;
}
template< typename T >istream& operator>>(istream& is, vector< T >& v) {
    for (T& in : v) is >> in;
    return is;
}
template<class T> bool chmax(T& a, const T& b) { if (a < b) { a = b; return 1; } return 0; }
template<class T> bool chmin(T& a, const T& b) { if (b < a) { a = b; return 1; } return 0; }
template <class T>
T div_floor(T a, T b) {
    if (b < 0) a *= -1, b *= -1;
    return a >= 0 ? a / b : (a + 1) / b - 1;
}
template <class T>
T div_ceil(T a, T b) {
    if (b < 0) a *= -1, b *= -1;
    return a > 0 ? (a - 1) / b + 1 : a / b;
}
template <class F> struct rec {
    F f;
    rec(F&& f_) : f(std::forward<F>(f_)) {}
    template <class... Args> auto operator()(Args &&... args) const {
        return f(*this, std::forward<Args>(args)...);
    }
};
lint gcd(lint a, lint b) { if (b == 0) return a; else return gcd(b, a % b); }
lint digit(lint a) { return (lint)log10(a); }
lint e_dist(plint a, plint b) { return abs(a.first - b.first) * abs(a.first - b.first) + abs(a.second - b.second) * abs(a.second - b.second); }
lint m_dist(plint a, plint b) { return abs(a.first - b.first) + abs(a.second - b.second); }
bool check_overflow(lint a, lint b, lint limit) { if (b == 0) return false; return a >= limit / b; } // a * b > c => true
void Worshall_Floyd(VVl& g) { REP(k, SZ(g)) REP(i, SZ(g)) REP(j, SZ(g)) chmin(g[i][j], g[i][k] + g[k][j]); }
const lint MOD1000000007 = 1000000007, MOD998244353 = 998244353, INF = 2e18;
lint dx[8] = { 0, -1, 0, 1, 1, -1, 1, -1 }, dy[8] = { -1, 0, 1, 0, -1, -1, 1, 1 };
bool YN(bool flag) { cout << (flag ? "YES" : "NO") << endk; return flag; } bool yn(bool flag) { cout << (flag ? "Yes" : "No") << endk; return flag; }
struct Edge {
    lint from, to;
    lint cost;
    Edge() {

    }
    Edge(lint u, lint v, lint c) {
        cost = c;
        from = u;
        to = v;
    }
    bool operator<(const Edge& e) const {
        return cost < e.cost;
    }
};
struct WeightedEdge {
    lint to;
    lint cost;
    WeightedEdge(lint v, lint c = 1) {
        to = v;
        cost = c;
    }
    bool operator<(const WeightedEdge& e) const {
        return cost < e.cost;
    }
};
using WeightedGraph = V<V<WeightedEdge>>;
typedef pair<plint, lint> tlint;
typedef pair<ld, ld> pld;
typedef pair<lint, tlint> qlint;
typedef pair<string, lint> valstr;
typedef pair<Vl, lint> valv;

template< typename G >
struct StronglyConnectedComponents {
    const G& g;
    VVl gg, rg;
    vector< int > comp, order, used;

    StronglyConnectedComponents(G& g) : g(g), gg(g.size()), rg(g.size()), comp(g.size(), -1), used(g.size()) {
        for (int i = 0; i < g.size(); i++) {
            for (auto e : g[i]) {
                gg[i].emplace_back((int)e);
                rg[(int)e].emplace_back(i);
            }
        }
    }

    int operator[](int k) {
        return comp[k];
    }

    void dfs(int idx) {
        if (used[idx]) return;
        used[idx] = true;
        for (int to : gg[idx]) dfs(to);
        order.push_back(idx);
    }

    void rdfs(int idx, int cnt) {
        if (comp[idx] != -1) return;
        comp[idx] = cnt;
        for (int to : rg[idx]) rdfs(to, cnt);
    }

    void build(VVl& t) {
        for (int i = 0; i < gg.size(); i++) dfs(i);
        reverse(begin(order), end(order));
        int ptr = 0;
        for (int i : order) if (comp[i] == -1) rdfs(i, ptr), ptr++;

        t.resize(ptr);
        for (int i = 0; i < g.size(); i++) {
            for (auto& to : g[i]) {
                int x = comp[i], y = comp[to];
                if (x == y) continue;
                t[x].push_back(y);
            }
        }
    }
};

struct UnionFind {
public:
    UnionFind() : _n(0) {}
    UnionFind(int n) : _n(n), parent_or_size(n, -1) {}

    int merge(int a, int b) {
        assert(0 <= a && a < _n);
        assert(0 <= b && b < _n);
        int x = leader(a), y = leader(b);
        if (x == y) return x;
        if (-parent_or_size[x] < -parent_or_size[y]) std::swap(x, y);
        if (used_count) {
            if (count_in_set[x].size() < count_in_set[y].size()) {
                std::swap(count_in_set[x], count_in_set[y]);
            }
            for (auto p : count_in_set[y]) {
                count_in_set[x][p.first] += p.second;
            }
        }
        if (set_operate) {
            root_values[x] = f(root_values[y], root_values[x]);
        }
        parent_or_size[x] += parent_or_size[y];
        parent_or_size[y] = x;

        return x;
    }

    bool same(int a, int b) {
        assert(0 <= a && a < _n);
        assert(0 <= b && b < _n);
        return leader(a) == leader(b);
    }

    int leader(int a) {
        assert(0 <= a && a < _n);
        if (parent_or_size[a] < 0) return a;
        return parent_or_size[a] = leader(parent_or_size[a]);
    }

    int size(int a) {
        assert(0 <= a && a < _n);
        return -parent_or_size[leader(a)];
    }

    std::vector<std::vector<int>> groups() {
        std::vector<int> leader_buf(_n), group_size(_n);
        for (int i = 0; i < _n; i++) {
            leader_buf[i] = leader(i);
            group_size[leader_buf[i]]++;
        }
        std::vector<std::vector<int>> result(_n);
        for (int i = 0; i < _n; i++) {
            result[i].reserve(group_size[i]);
        }
        for (int i = 0; i < _n; i++) {
            result[leader_buf[i]].push_back(i);
        }
        result.erase(
            std::remove_if(result.begin(), result.end(),
                [&](const std::vector<int>& v) { return v.empty(); }),
            result.end());
        return result;
    }
    //update root calc
    //set by set operations
    void set_operate_and_value(std::vector<lint> array, function<lint(lint, lint)> _f) {
        f = _f;
        root_values = array;
        set_operate = true;
    }
    lint get_set_value(int a) {
        return root_values[leader(a)];
    }

    //regist count
    void regist_count(int a, int label) {
        if (!used_count) {
            used_count = true;
            count_in_set.assign(_n, std::map<int, int>());
        }
        count_in_set[leader(a)][label]++;
    }

    int get_count(int a, int label) {
        if (!used_count) return -1;
        return count_in_set[leader(a)][label];
    }

private:
    int _n;
    std::vector<int> parent_or_size;
    std::vector<std::map<int, int>> count_in_set;
    bool used_count = false;
    std::vector<lint> root_values;
    function<lint(lint, lint)> f;
    bool set_operate = false;
};

int main() {
	lint N, M;
    cin >> N >> M;
    VVl to(N);
    REP(i, M) {
        lint u, v;
        cin >> u >> v; u--; v--;
        to[u].push_back(v);
    }
    StronglyConnectedComponents g(to);
    VVl res;
    g.build(res);
    set<lint> st;
    REP(i, N) st.insert(g[i]);
    V<plint> ed;
    UnionFind tree(SZ(res));
    REP(i, SZ(res)) {
        for (lint v : res[i]) {
            ed.push_back({ i, v });
            tree.merge(i, v);
        }
    }
    map<lint, V<plint>> med;
    REP(i, SZ(ed)) {
        med[tree.leader(ed[i].first)].push_back(ed[i]);
    }
    lint ans = 0;
    set<lint> used;
    VVl vers(SZ(res));
    REP(i, SZ(res)) vers[tree.leader(i)].push_back(i);
    REP(i, SZ(res)) {
        lint _p = tree.leader(i);
        if (used.count(_p)) continue;
        used.insert(_p);

        lint sz = SZ(vers[_p]);
        if (sz == 1) continue;
        sort(ALL(vers[_p]));
        map<lint, lint> fx;
        for (lint v : vers[_p]) fx[v] = SZ(fx);
        VVl nxt(sz);
        Vl deg(sz);
        for (plint p : med[_p]) {
            nxt[fx[p.first]].push_back(fx[p.second]);
            deg[fx[p.second]]++;
        }
        queue<lint> que;
        REP(j, sz) if (deg[j] == 0) que.push(j);
        Vl res;
        while (!que.empty()) {
            lint curr = que.front(); que.pop();
            res.push_back(curr);
            for (lint v : nxt[curr]) {
                deg[v]--;
                if (deg[v] == 0) que.push(v);
            }
        }
        Vl dp(sz);
        for (lint v : res) {
            for (lint _n : nxt[v]) {
                chmax(dp[_n], dp[v] + 1);
            }
        }
        ans += sz - *max_element(ALL(dp)) - 1;
    }
    ans += SZ(used);
    if (SZ(res) == 1) ans = 0;
    cout << ans << endk;
}